industry move that matters for local users for Llama

[southpatch]

Interesting, I hadn't tried that. I ended up in about the same place using PyTorch. The biggest change for me was simplifying the stack instead of adding one more optimization layer.